Automatic identification of anatomical landmarks in three-dimensional computed tomography/cone-beam computed tomography: a scoping review.
Authors
Affiliations (4)
Affiliations (4)
- Department of Stomatology, The First Affiliated Hospital of Fujian Medical University, Fuzhou, China.
- School of Stomatology, Fujian Medical University, Fuzhou, China.
- School of Basic Medical Sciences, Fujian Medical University, Fuzhou, China.
- Department of Stomatology, National Regional Medical Centre, Binhai Campus of the First Affiliated Hospital, Fujian Medical University, Fuzhou, China.
Abstract
This study aimed to conduct a scoping review to systematically review automatic identification techniques for soft-tissue and hard-tissue landmarks in three-dimensional (3D)-computed tomography (CT)/cone-beam computed tomography (CBCT), particularly focusing on artificial intelligence (AI)-based methods, to explore the progress and challenges in accuracy, efficiency, and clinical applicability. In this scoping review, we searched for studies on automatic landmarking in CT/multi-slice spiral CT or CBCT that were published until January 2026 in PubMed, Web of Science, and the Cochrane Library. Studies that validated automatic landmarking methods against a reference standard, typically defined by manual annotations from human experts, were included. In total, 37 (20 CBCT-only, 10 CT-only, and 7 CBCT + CT) studies were included in this review. Most of the studies focused on non-syndromic permanent dentition populations or a single type of malocclusion (except for 6 that involved mixed dentition) and had a limited number of samples, with <200 cases included in 75.7% of the studies. With 7-105 landmark annotations, most studies focused on hard-tissue marker points, and only 6 involved soft-tissue landmarks. Regarding accuracy, 56.8% of the studies met the clinical correctness standard [mean radial error (MRE) < 2 mm], and the overall trend indicated a gradual increase in the landmarking accuracy. Some studies did not use MRE or successful detection rate as an outcome measure, which potentially affected the overall comparability of the analysis. Existing studies have limited samples (number and type); the landmarks are mostly focused on hard tissues, and single algorithms are limited in their robustness and generalization performance in clinical applications. Furthermore, the system for evaluating the accuracy of 3D-automatic landmarking has not been standardized, and the clinical significance of traditional two-dimensional-accuracy thresholds in the 3D-space remains controversial.